.. _digital-intermediate:
-VA-API H.264 encoding
----------------------
+VA-API H.264 encoding (optional)
+--------------------------------
Even on modern networks and with today's large SSDs, uncompressed HD video is a
bit unwieldy to send around (uncompressed 720p60 4:2:0 is about 79 MB/sec,
(for future editing or re-streaming) or sent to an encoder on another machine
for final streaming.
-Currently, only VA-API is supported for encoding the digital intermediate, although Nageru might
-support NVIDIA's NVENC at some point in the future. In particular, this means that Intel Quick
-Sync Video (QSV), the hardware H.264 encoder present on all modern Intel
-GPUs, is supported. QSV is more than fast enough to keep up with 720p60 in
+Although you can (since Nageru 1.5.0) use software encoding through x264 for
+the digital intermediate, it is generally preferred to use a hardware encoder
+if it is available. Currently, VA-API is the only hardware encoding method
+supported for encoding the digital intermediate, although Nageru might support
+NVIDIA's NVENC at some point in the future. In particular, this means that
+Intel Quick Sync Video (QSV), the hardware H.264 encoder present on all modern
+Intel GPUs, is supported.
+
+QSV is more than fast enough to keep up with 720p60 in
realtime without eating appreciably into the power budget, but it is not
competitive with the top H.264 encoders in terms of quality per bit. Also,
the stream is encoded using *constant quality* (fixed quantizer), not
Frame rates are automatically converted; one input is designated as the
**master clock** (right-click on an input to select it as such), and gets
to dictate the frame rate of the output. Inputs with differing frame rates
-will get frames duplicated or dropped as needed (with adaptive queueing to
+will get frames duplicated or dropped as needed (with adaptive queuing to
account for clock and jitter).
Nageru works in 16-bit floating-point RGBA internally. High-quality conversion to and
from subsampled Y'CbCr (typically 4:2:2 for inputs and 4:2:0 for outputs)
-is done transparently on the GPU.
+is done transparently on the GPU. Input and output is 8-bit Y'CbCr by default,
+but be aware that 8-bit Y'CbCr, however common, cannot capture the full color
+fidelity of 8-bit RGB (not to mention 10-bit RGB). If you have spare GPU power,
+you can enable 10-bit Y'CbCr input and output with --10-bit-input and
+--10-bit-output, respectively, although you should be aware that client
+support for 10-bit H.264 is very limited. Also, Quick Sync Video does not
+support 10-bit H.264 encoding, so in this case, the digital intermediate needs
+to be encoded in software.
+
+
+Performance tips
+----------------
+
+It is strongly recommended to have the rights to run at real-time priority;
+it will make the USB3 threads do so, which will make them a lot more stable.
+(A reasonable hack for testing is probably just to run it as root using sudo,
+although you might not want to do that in production, but instead grant
+your regular user permissions in /etc/security/limits.conf.) Note also that if you
+are running a desktop compositor, it will steal significant amounts of GPU
+performance. The same goes for PulseAudio.
+
+Nageru tries to lock itself into RAM if it has the permissions to do
+so, for better realtime behavior. (Writing the stream to disk tends to
+fill the buffer cache, eventually paging less-used parts of Nageru out.)
+Again, this is something you can set in limits.conf.
projects / nageru-docs / blobdiff